The Faithfulness Gap: Certifying Semantic Equivalence Between Natural-Language and Formal Mathematical Statements
Researchers have introduced Bidirectional Provability Fingerprinting (BPF), a new framework designed to certify the faithfulness of autoformalized mathematical statements. This method addresses the challenge where translated formal statements may be provable but not semantically equivalent to the original natural-language intent. The framework includes components for generating counterfactual probes, an equivalence spectrum for continuous scoring, adaptive budget allocation, and faithfulness-guided decoding. A new benchmark, DriftBench, comprising 2,183 NL/Lean 4 pairs, was also released to evaluate these methods. AI
IMPACT This research aims to improve the reliability of AI systems translating natural language mathematics into formal proofs, potentially increasing trust in AI-assisted mathematical discovery.